JPH0438587A - Input area adaptive type neural network character recognizing device - Google Patents

Abstract

PURPOSE:To improve character recognizing accuracy and sureness by executing character recognition with a neural network adaptively to a character pattern including noise or with a input area part turned or parallel-moved to the character pattern cut out roughly. CONSTITUTION:A standard character pattern category memory part 11 holds N pairs of a standard character pattern for learning and the category of it. A pattern associative neural network 12 makes the standard character pattern as an input signal and learns with the category of the character pattern as an educator signal and possesses a function to recognize a category K1 of the character pattern. An input area part 14 includes the respective input cells of the pattern associative neural network 12 and the pattern from an external part becomes the input of the pattern associative neural network 12 through the input area part 14.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a device for recognizing classification of patterns, and more particularly to a device for recognizing character patterns containing noise or rotated or translated.

[Conventional technology]

Conventionally, many character recognition systems capable of handling classification problems have been proposed. Regarding these, for example, "Pattern Recognition", supervised by Takeshi Mori, published by the Institute of Information and Communication Engineers, (1
988) (hereinafter referred to as cited document 1). The process of pattern recognition consists of preprocessing, feature extraction, and identification of the original pattern signal. In the pre-processing process,
Scale the size of the pattern so that it is the same size as the pattern you learned. In the process of feature extraction, properties of patterns that are effective for recognition are extracted as features. When a character pattern that contains noise or is rotated or translated in parallel is presented, the preprocessing or feature extraction process is not performed properly, resulting in incorrect recognition.

[Problem to be solved by the invention]

In the above-mentioned conventional technology, in order to achieve a sufficient recognition rate, it is necessary to appropriately perform a preprocessing or feature extraction process. However, it is difficult to predetermine appropriate preprocessing or feature extraction steps for all possible patterns.

An object of the present invention is to improve both character recognition precision and accuracy by performing character recognition using a neural network adaptively to character patterns that include noise or whose input area has been rotated or translated. An object of the present invention is to provide an input area adaptive neural network character recognition device that can perform the following functions.

[Means to solve the problem]

The input domain adaptive neural network character recognition device of the present invention includes a standard character pattern/category memory unit that holds standard character patterns and categories of the character patterns, and a character pattern that is rotated or translated and contains noise. an input area that adapts to the input area, a pattern associative neural network that can recognize character patterns presented in standard positions, and a pattern associative neural network that controls the pattern associative neural network. The present invention is characterized by having a learning/recognition control section that performs the following functions.

[Effect]

In the present invention, a character pattern roughly cut out using a conventional method is used as input. Character recognition is performed on this pattern using a pattern associative neural network. The recognition result is used to adapt the input area to the character pattern. Character recognition is performed again using the neural network on the adapted input. This procedure is repeated to recognize characters. Preprocessing is performed sequentially based on the recognition results, and is performed appropriately even for rotated or translated character patterns that include noise. Further, since the feature extraction process is performed by bilateral network learning, appropriate feature extraction is performed for the character pattern to be classified.

The pattern associative neural network used here includes, for example, "Nikkei Electronics" magazine issue 427 (19
(August 1987), pages 15 to 124 [Neural
A pattern-associative neural network, which is explained in detail in the article entitled "Using the Internet for Pattern Recognition, Signal Processing, and Knowledge Processing" (hereinafter referred to as Cited Document 2), can be used.

FIG. 6 is a structural explanatory diagram showing an example of the structure of a pattern associative neural network. This pattern associative neural network has an input layer 51, an intermediate layer 52. Each layer of the output layer 53 has a hierarchical structure. Note that the middle layer 52
Although it has one layer in FIG. 6, it may have multiple layers of two or more layers.

The output of a node in each layer of a pattern associative neural network is given a weight W to the nodes connected to that node.
The value of the sum of the products multiplied by is converted by a nonlinear function. In this way, the conversion characteristics of the pattern associative neural network are determined by the weights W. The value of weight W is determined by learning. The learning method can be implemented using, for example, backward propagation, which is explained in detail in Cited Document 2.

Assume that an input character pattern P is presented to a neural network that has completed learning. From now on, input layer 5
The value of each cell of 1 is determined. The output value of each cell of the output layer 53 is determined using forward propagation. Since each cell in the output layer 53 is associated with a character category, the cell with the largest output value becomes the category of the input character pattern P.

When a character pattern containing noise or rotated or translated is presented, the output of the cell of the output layer 53 corresponding to the category of the input character pattern P becomes maximum due to the generalization ability of the neural network.

However, outputs also appear in cells of the output layer 53 that do not correspond to the input character pattern P because they contain noise or are rotated or translated.

At this time, the value that each cell of the output layer 53 should output is 1.0 for cells in the corresponding category and 0.0 for cells in the non-corresponding category. Calculate the error of the neural network using the value to be output and the output value. here,
Backward and detailed explanations in cited document 2
The δ value of the cell in the input layer 51 is determined using propagation.

The δ value is: where E is the error described above and net is the input of the cell of the input layer 51.

Since the δ value is the partial differential coefficient of the input with respect to the error, if the input of the cell of the input layer 51 is changed by one ηδ (however, 0〈
η<1), the error is reduced.

The input of a cell in the input layer 51 is changed by moving the cell using the local pattern density gradient of the cell in the input region. At the same time, constraints are placed between cells in the input area so that they do not move chaotically.

By moving the cells in the input area, the value of each cell in the input layer 51 is newly determined. The output value of each cell of the output layer 53 is determined using forward propagation.

By repeating the above procedure, the input domain is adapted to reduce errors in the output layer.

That is, the input area is adapted so that the character pattern becomes the same as the learned character pattern, and appropriate preprocessing is performed even for character patterns that include noise or are rotated or translated. Furthermore, since the feature extraction process is performed by neural network learning, appropriate feature extraction is performed for the character pattern to be classified.

〔Example〕

Embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the human-power domain adaptive neural network character recognition device of the present invention. This input domain adaptive neural network character recognition device is
Standard character pattern/category/(division) memory section 1
1, a pattern associative neural network 12,
Learning recognition control unit 13, human power area unit 14, and switching unit 1
5.

The standard character pattern/category/memory unit 11 holds N sets of standard character patterns for learning and their categories (classifications).

The pattern associative neural network 12 uses the standard character pattern St as an input signal, learns the category of that character pattern as a teacher signal, and creates the character pattern S.
t has a function of recognizing its category Kt.

The learning/recognition control unit 13 controls learning and recognition operations of the pattern associative neural network 12.

The input area section 14 includes each input cell of the pattern associative neural network 12, and a pattern from the outside becomes an input to the pattern associative neural network 12 through the input area section 14.

The switching unit 15 switches the input to the pattern associative neural network 12 between the learning and recognition phases.

Next, the operation of this embodiment will be explained.

The operation consists of a learning phase in which the pattern associative neural network 12 is trained, and a recognition phase in which character patterns are recognized using the device.

(1) Learning phase The learning phase will be explained using FIG. 1 and FIG. 2 showing an example of a signal flow in the learning phase.

In the learning phase, the switching unit 15 in FIG. 1 is switched to the contact a side.

The pattern associative neural network 12 uses the standard character pattern S1 from the standard character pattern/category/memory section 11 as an input signal, and uses the category Kt of that character pattern as a teacher signal.

The internal parameters of the pattern associative neural network 12 are updated based on these input signals and teacher signals, and this update is performed using backward propagation, which is explained in detail in Cited Document 2. The learning operations described above are performed on the standard character pattern S1 and the category Kt held in the standard character pattern/category/memory section 11. If the error after learning is not small enough, the learning operation described above is repeated until the error becomes small enough.

This completes the learning of the pattern associative neural network 12. Through this learning, the pattern associative neural network 12 has a function of recognizing the category Kt based on the character pattern SL through the aforementioned learning.

(2) Learning phase of recognition Regarding the recognition phase, see Figure 1, Figure 3 showing an example of the signal flow in the recognition phase, Figure 4 showing the procedure in the recognition phase, and the gradient G of pattern density. This will be explained with reference to FIG. 5, which is an explanatory diagram of the calculation.

In the recognition phase, the switching section 15 in FIG. 1 is switched to the contact side.

First, a pattern to be recognized is presented to the input area section 14 from the outside. The pattern to be recognized at this time is cut out more roughly than the original pattern. This technique is described in reference l.

The calculation of the input of each input cell of the pattern associative neural network 12 (see the step knob 41 in FIG. 4) is performed according to the strength of the turn presented at the position of the input area. The input (ΔX, Δy) of a cell located at a position shifted by ΔX, Δy from the grid point can be determined as follows, for example.

l (ΔX, Δy) = (1-ΔX) (1-Δy)Is
.

+Δx(1−Δy)It.

+ (1-ΔX)ΔyIoI +ΔXΔ)11++ (2) Here, t...
1. . ,■. l+I. . is the density of the pattern corresponding to each grid point.

The calculation of the output (step 42 in FIG. 4) can then be performed using forward propagation, which is explained in detail in Reference 2.

When selecting the "correct answer" (Step 43 in FIG. 4), the one with the maximum output among the outputs is selected as the correct answer. Once the correct answer is determined, the mean squared error E will be as follows.

In the formula, N is the number of cells in the output layer, and A is 1.0 if A has the maximum output, and 0.0 otherwise.

0 is the output of the i-th cell of the output layer.

Calculating the delta value for each cell in the input layer (step 44 in Figure 4)
) can be performed using backward propagation, which is explained in detail in Reference 2. With this, the delta value δ of the j-th cell of the input layer can be determined. The delta value δ has the following relationship with the mean square error E in the output layer.

Here, neJ is the input of the j-th input cassette.

Therefore, when δ is a positive value, as the total input neJ of the j-th input cassette increases, the mean square error E decreases. As a result, the output will be closer to the "correct answer". From this, the output becomes more “correct”.

It can be seen that in order to approach Δnet, the total input netJ of the j-th input cassette needs to be changed by Δnet as given below.

Δnet, -ηδ, (5)
“The position that must be moved” to reduce the error in the output layer is the gradient G of the pattern density and Δnetj
Calculate using.

The gradient G of the pattern density can be calculated by interpolation as shown below based on the surrounding pattern density, as shown in FIG. 5, for example.

Gj = c, , Ji +cy, aj
(6) = Δ)' (Iz Ion) + (1-Δy) (1+.-10°) (7) = ΔX
(Iz 1111) + (1-Δx) (Ion Ioo)
(8) Here, Iz, It. ,Io,,I. . is the density of the pattern corresponding to each position.

The force that attempts to move cells in order to reduce output errors is hereinafter referred to as soltaka.

Soltaka n, is ΔneJ and the pattern density gradient G
, it can be calculated as follows.

In the above equation, when the gradient of pattern density cyj=o, the number n becomes infinite. However, considering the above fact that in this state the soltaka nj must be approximately 0, the soltaka Qa can be determined as follows, for example.

The spatial constraint that maintains a rectangular shape between cells in the input area is hereinafter referred to as address force. Addressing force occurs when the distance between cells increases and attempts to return to the original distance.

This force works on keeping the input area rectangular,
Addressing force FJ+@, which does not work even if the entire input area is rotated or translated, and which acts on the j-th cell and the m-th cell, is given as follows.

22, k is a spring constant, JM is a position vector whose starting point is given by the j-th cell and whose ending point is given by the m-th cell, and 41. teeth,
This is the natural length of the j-th cell and the m-th cell.

Using the address force pj+a acting on the j-th cell and the m-th cell, the address force FJ+Il acting on the j-th cell is as follows.

FJ=ΣFj,, 021
Here, determining the new position of the cell in the input area (step 45 in FIG. 4) determines the new position of the cell by combining the soltaka and address force. The new position vector of the cell, P, can be calculated as follows.

P, *mw = P, old ΔPJ"'
Q31ΔP j"' = α, ΔP J"'+
a! o40α, F j 04)
α1 is a parameter that affects the amount of change in the previous position,
It can be considered a kind of inertia. By using this inertia, it is possible to overcome the local minimum and adapt the input region to the pattern.

If the error has not converged to a sufficiently small value,
If the input of each input cell is converged to the calculation (step 41 in FIG. 4), the process ends.

[Effects of the Invention] The input area adaptive neural network character recognition device of the present invention is adaptive to character patterns that are roughly cut out, contain noise, or whose input area portions have been rotated or translated in parallel. By performing character recognition using a neural network, both character recognition precision and accuracy can be improved. Moreover, it has the effect that the presentation position of the pattern can be determined.

[Brief explanation of drawings]

Figure 1 is a block diagram showing an example of an input domain adaptive neural network character recognition device, Figure 2 is a block diagram showing the signal flow in the learning phase, and Figure 3 is a block diagram showing the signal flow in the recognition phase. 4 is a flowchart showing the procedure in the recognition phase, FIG. 5 is an explanatory diagram for determining the pattern density gradient Gj, and FIG. 6 is a diagram showing an example of the configuration of a pattern associative neural network. 11...Standard character pattern/category/memory unit 12...Pattern associative neural network 13...Learning/recognition control unit 14...Input area unit 15...・Switching unit 51...Input layer 52...Middle layer 53...Output layer

Claims (1)

[Claims] (1) A standard character pattern category memory section that holds standard character patterns and the categories of the character patterns; an input area section that adapts to rotated or translated character patterns that include noise; An input area adaptive neural network recognition device comprising: a pattern associative neural network capable of recognizing a character pattern presented at a specific position; and a learning/recognition control unit that controls the pattern associative neural network. .